From: Lorenz Bauer <lmb@xxxxxxxxxxxxx> Date: Wed, 9 Aug 2023 16:08:31 +0100 > On Wed, Aug 9, 2023 at 3:39 PM Martin KaFai Lau <martin.lau@xxxxxxxxx> wrote: > > > > On 8/9/23 1:33 AM, Lorenz Bauer wrote: > > > Kumar reported a KASAN splat in tcp_v6_rcv: > > > > > > bash-5.2# ./test_progs -t btf_skc_cls_ingress > > > ... > > > [ 51.810085] BUG: KASAN: slab-out-of-bounds in tcp_v6_rcv+0x2d7d/0x3440 > > > [ 51.810458] Read of size 2 at addr ffff8881053f038c by task test_progs/226 > > > > > > The problem is that inet[6]_steal_sock accesses sk->sk_protocol without > > > accounting for request sockets. I added the check to ensure that we only > > > every try to perform a reuseport lookup on a supported socket. > > > > > > It turns out that this isn't necessary at all. struct sock_common contains > > > a skc_reuseport flag which indicates whether a socket is part of a > > > > Does it go back to the earlier discussion > > (https://lore.kernel.org/bpf/7188429a-c380-14c8-57bb-9d05d3ba4e5e@xxxxxxxxx/) > > that the sk->sk_reuseport is 1 from sk_clone for TCP_ESTABLISHED? It works > > because there is sk->sk_reuseport"_cb" check going deeper into > > reuseport_select_sock() but there is an extra inet6_ehashfn for all TCP_ESTABLISHED. > > Sigh, I'd forgotten about this... > > For the TPROXY TCP replacement use case we sk_assign the SYN to the > listener, which creates the reqsk. We can let follow up packets pass > without sk_assign since they will match the reqsk and convert to a > fullsock via the usual route. At least that is what the test does. I'm > not even sure what it means to redirect a random packet into an > established TCP socket TBH. It'd probably be dropped? > > For UDP, I'm not sure whether we even get into this situation? Doesn't > seem like UDP sockets are cloned from each other, so we also shouldn't > end up with a reuseport flag set erroneously. > > Things we could do if necessary: > 1. Reset the flag in inet_csk_clone_lock like we do for SOCK_RCU_FREE I think we can't do this as sk_reuseport is inherited to twsk and used in inet_bind_conflict(). > 2. Duplicate the cb check into inet[6]_steal_sock or 3. Add sk_fullsock() test ?
